A late fusion multi-task learning for respiratory waveform and rate estimation from photoplethysmography
Clinical Snapshot
PICO Framework
| P — Population | Adult patients with continuous physiological monitoring data available from two benchmark datasets: CapnoBase (n=42) and BIDMC (n=52) |
| I — Intervention | Late fusion multi-task deep learning framework processing three PPG-derived respiratory modulations (RIIV, RIAV, RIFV) with dual decoders for simultaneous respiratory waveform reconstruction and respiratory rate estimation |
| C — Comparator | Single-task deep learning architectures; alternative training strategies including pooled training and non-transfer-learning configurations; implicitly compared against conventional capnography as reference standard |
| O — Outcomes | Primary: Mean absolute error (MAE) for respiratory rate estimation (bpm) and respiratory waveform reconstruction (%). Secondary: Pearson correlation coefficient (r) for waveform fidelity; comparative performance across training strategies (sequential transfer learning, pooled training, dataset-specific training) |
Bottom Line
This paper presents a technically novel late fusion multi-task deep learning framework that simultaneously estimates respiratory rate and reconstructs respiratory waveforms from PPG-derived modulations. The key methodological contribution is demonstrating that joint multi-task training outperforms single-task approaches, and that sequential transfer learning is superior to pooled training across incompatible reference standards — a practically important finding for dataset curation in this field. Respiratory rate estimation performance on the BIDMC dataset (MAE 1.33 bpm) is clinically competitive, though the CapnoBase result (2.27 bpm) is borderline for high-acuity applications. Waveform reconstruction accuracy is moderate (r=0.59–0.66) and its clinical utility at this fidelity level remains unestablished. Critical limitations include very small benchmark datasets (n=42, n=52), absence of confidence intervals, no prospective clinical validation, and no evaluation of clinically meaningful outcomes such as detection sensitivity for respiratory deterioration. The study represents a credible proof-of-concept contribution to the PPG respiratory monitoring literature but is several validation steps removed from clinical deployment. Senior clinicians should regard this as early-phase signal processing research requiring prospective clinical validation before any consideration of implementation.
Key Findings
P Value: Not reported in abstract
Effect Size: Respiratory rate MAE: 2.27 bpm (CapnoBase), 1.33 bpm (BIDMC); Waveform reconstruction MAE: 19.00% (CapnoBase), 20.90% (BIDMC); Waveform Pearson r: 0.662 (CapnoBase), 0.591 (BIDMC)
Primary Outcome: Simultaneous respiratory rate estimation and respiratory waveform reconstruction from PPG using a late fusion multi-task deep learning framework
Nnt Or Sensitivity: Sensitivity/specificity for clinically relevant respiratory rate thresholds not reported; diagnostic performance metrics absent. Sequential transfer learning consistently outperformed pooled and non-transfer strategies across both datasets and both tasks.
Confidence Interval: Not reported
Clinical Application
PPG is widely available via pulse oximeters already deployed in most clinical settings, making hardware implementation feasible without additional patient burden. However, the proposed framework requires software integration into monitoring systems, and real-time computational feasibility has not been demonstrated. Waveform reconstruction accuracy at current performance levels (r≈0.59–0.66) may be insufficient for clinical decisions dependent on waveform morphology. Respiratory rate estimation performance (MAE 1.33–2.27 bpm) approaches but does not consistently meet the ≤2 bpm threshold generally considered acceptable for high-acuity monitoring. In Australia, continuous respiratory monitoring is a recognised priority in the National Safety and Quality Health Service Standards, particularly for patients receiving opioid analgesia and those at risk of deterioration on general wards. The RACGP and ACSQHC have highlighted respiratory rate as an underutilised vital sign. A validated, non-invasive PPG-based respiratory monitoring solution would have direct relevance to Australian ward-based monitoring programs and could complement existing MET (Medical Emergency Team) activation criteria. However, TGA regulatory approval as a Class IIb or III medical device would be required before clinical deployment. PBS listing is not applicable to this technology category. The framework has not been validated in Australian patient cohorts, and performance in populations with higher rates of obesity, Indigenous Australians, or patients with chronic respiratory disease remains unknown. Adult patients undergoing continuous physiological monitoring where non-invasive respiratory monitoring is clinically indicated — including post-operative recovery, general ward monitoring, sleep studies, and remote patient monitoring. Not yet validated for paediatric populations, critically ill patients with haemodynamic instability, or patients with conditions significantly affecting PPG signal quality.
Abstract
Continuous respiratory monitoring enables early detection of physiological deterioration, yet conventional capnography remains impractical for prolonged use. Photoplethysmography (PPG) offers a non-invasive alternative that encodes respiratory information through baseline wander (respiratory-induced intensity variation; RIIV), amplitude modulation (respiratory-induced amplitude variation; RIAV), and frequency modulation (respiratory-induced frequency variation; RIFV) of the pulsatile waveform. Existing PPG-based deep learning approaches, whether operating on the raw signal or on these physiological modulations, are limited to single-task architectures that estimate either respiratory rate or reconstruct the respiratory waveform in isolation, without jointly addressing both outputs. We propose a late fusion multi-task framework in which dedicated encoder branches independently process each modulation before fusion, and dual decoders simultaneously reconstruct the respiratory waveform and estimate the respiratory rate. The framework was evaluated on the CapnoBase (n = 42) and BIDMC (n = 52) benchmarks across multiple training strategies. For respiratory-rate estimation, the best transfer-learning configurations achieved a mean absolute error (MAE) of 2.27 bpm on CapnoBase and 1.33 bpm on BIDMC. For waveform reconstruction, the corresponding MAE values were 19.00% and 20.90%, with moderate correlations (r = 0.662 and r = 0.591, respectively). Sequential transfer learning consistently outperformed all other strategies, whereas pooled training degraded both outputs, demonstrating that capnography-derived and impedance-derived waveforms are not interchangeable training targets. These findings establish that short-window PPG can simultaneously support respiratory-rate estimation and waveform reconstruction, when reference signal compatibility is explicitly addressed in multi-task training.
References
- 1.Ho, M. N., & Nguyen, K. T. (2026). A late fusion multi-task learning for respiratory waveform and rate estimation from photoplethysmography. PLoS ONE. https://doi.org/10.1371/journal.pone.0353203
Related Research
Sleep medicine
Concordance of wearable device sleep metrics with patient-reported sleep quality: A systematic review
3 Aug 2026
Bioelectrochemistry (Amsterdam, Netherlands)
From data to diagnosis: A comprehensive review of machine learning-driven wearable sensors in healthcare.
2 Aug 2026
JMIR mHealth and uHealth
Wearable Devices for Monitoring and Management of Comorbid Obstructive Sleep Apnea and Hypertension: Scoping Review
1 Aug 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service